Google Threat Intelligence Group and Mandiant reported the first observed case of attackers using an autonomous AI framework to turn newly released security patches into working exploits with no human in the loop. Researchers found a publicly exposed staging server running a multi-agent framework called exploit_agent that connected frontier and locally hosted models to automate patch diffing, exploit generation and execution. The server also held stolen victim data and post-compromise scripts for reconnaissance, brute-force attacks and webshells.
Google released Gemini 4 Hydrogen, the successor to its Flash models, in Search, the Gemini app, AI Studio, the Gemini API and Antigravity, and GitHub made it generally available in Copilot. Google says it is 2.5 times faster than Gemini 3.8 Flash in high mode, cuts fabricated claims by more than 5x and more than triples its Terminal-Bench score. A new Antigravity feature, Orchestrator, pairs Hydrogen with Gemini 4 Carbon, using 24% less quota and finishing tasks 19% faster, per Google.
Alibaba has formally unveiled Qwen4-Max, the first model in its next-generation Qwen4 family, aimed at complex professional tasks and autonomous AI agents. The model uses a new mixture-of-experts architecture with native vision and a 1M-token context window, and Alibaba says it outperforms Qwen3.8-Max while using up to about 50% fewer tokens. It was teased last week and offered in preview through several Alibaba products. Unlike earlier open-weight Qwen releases, Alibaba appears to be keeping its most powerful model proprietary.
Google released Gemini 4 Carbon, a frontier model for complex software engineering, cybersecurity and agentic work, and it is now generally available in GitHub Copilot for Pro+, Max, Business and Enterprise users. Google DeepMind also said its multi-agent AlphaProof Code framework formally verified Signal's PQXDH handshake and Triple Ratchet in Lean, a protocol also used by WhatsApp and Google Messages. Google plans to open-source AlphaProof Code. The release adds /swarm, a preview Antigravity feature coordinating multiple coding agents.
CISA confirmed active exploitation of the two PHP2Shell vulnerabilities, CVE-2026-109618 and CVE-2026-109619, by adding them to its Known Exploited Vulnerabilities catalog. Both flaws carry a CVSS score of 9.2 and can enable unauthenticated remote code execution on affected PHP servers. Federal civilian agencies must patch by October 29, while researchers have also documented successful exploitation. Patches were released on October 21, and attackers reportedly attempted exploitation before public disclosure.
Jonathan Bryce, executive director of the Cloud Native Computing Foundation, said at KubeCon + CloudNativeCon North America in Salt Lake City that many existing systems were built for human speed and are not scalable or resilient enough for the workloads AI agents generate. He argued that financial services, healthcare and transportation systems that agents interact with will need the elasticity and observability of cloud-native technologies. At the event, Nvidia announced it is contributing OpenShell to the CNCF, and Google also announced a donation.
DuckDB's developers released DuckDB 2.0, codenamed Cinnamon Teal, in October, the analytical database's biggest feature release more than two years after 1.0.0. Storage format changes mean older versions may be unable to read files written by the new one. The release can run as a server, queries data faster, enables an extensible SQL parser by default and ships the Quack remote protocol as a first-class feature. DuckLabs' Gábor Szárnyas urges caution before moving production workloads to 2.0 as the team works on 2.0.1.
OpenCode developer Dax Raad rebuilt the project's browser automation tool, OpenCode Browse, in Rust for $60, replacing the parts of Playwright and Patchright it used. The new binary is 4.3 MB versus 53 to 64 MB for the Node version, and memory per connector fell from roughly 190 MB to 5.6 MB, about 97% less. The relay linking AI coding agents to the browser now uses 6 MB idle, against 353 MB for the old Node relay after six days. Browse connects agents to a developer's browser via a Chromium extension and an MCP server.
DuckDB's Quack is an HTTP-based protocol that streams query results in batches of columnar data as they are produced, letting developers work with results while a query still runs. It was previewed in May and released in October as part of DuckDB 2.0.0 LTS. DuckLabs product lead Guillermo Sánchez Dionis says older row-by-row protocols repeat column names and types with every row, creating bloated messages for large results.
Anthropic released Claude Fable 5.2 on November 9, saying it beats Opus 5.5 on most measures while recommending Opus as the default for coding and knowledge work. Anthropic reports Fable at 49.1% on FrontierCode 2.0 versus 45.5% for Opus, and 69.9% versus 62.3% on FrontierSWE v2. Fable costs $8/$40 per million input/output tokens against Opus's $4/$20; small hands-on coding and agent tests found similar results at almost four times the cost.
Amazon Web Services released AWS Emulator, a local copy of AWS that coding agents can test against. It covers more than 200 services and was written entirely by agents. Swami Sivasubramanian described how AWS's Kiro agent spent four nights failing to rebuild the replication engine of Alf, a Paxos-based lock service behind S3 and DynamoDB, but finished in about two hours once given a spec and test environment. The author argues that agents depend on verification loops and that realism is the hard part.
Anthropic says Claude Fable 5.2 triggered about 55% fewer safeguard blocks than Fable 5.1 overall and about 60% fewer in cybersecurity, and generates output up to 30% faster. API prices drop 20% to $8 per million input tokens and $40 per million output tokens, from $10 and $50. A hands-on comparison on three developer tasks, run once per model, found the gains played out unevenly, with only small savings on a login fix.
AI agents without search tools often invent URLs for current information, hit missing pages, and fall back on outdated training data. In October, Cloudflare opened a beta Web Search API in its AI Gateway that lets agents query several search providers through one interface, returning live results with source links, logging and access controls. Live search is likely to become a default agent tool, but deciding which fresh pages and internal documents to trust remains a ranking problem.
DuckLabs released DuckLake 1.1 in October alongside DuckDB 2.0.0 LTS. DuckLake is an open-source lakehouse table format that stores table metadata in a standard SQL database rather than in manifest files spread across object storage, with data served from Parquet files. Lead developer Pedro Holanda said such formats exist so users can switch database systems without migrations, and DuckLake treats lakehouse metadata as a relational database problem to cut complexity.
Anthropic's Tuesday expansion of Claude's enterprise lineup has raised concern among founders that their model supplier could become a direct competitor. GlacierGrid COO Gerald Zingraf compared the strategy to Amazon's private-label business, saying it makes him nervous for wrappers and pure SaaS firms, and suggested Anthropic can see which integrations customers use most, though he offered no evidence it uses that data. Other developers, including legal tech CEO Scott Stevenson, say the risk is already real.
Anthropic launched Claude Fable 5.2 on Nov. 9, reporting wins over OpenAI's GPT-6 Astra of 49.1% to 42.8% on FrontierCode 2.0 and 69.9% to 65.5% on FrontierSWE v2, though the 49.1% came at X high effort versus 45.3% at the default High. Fable 5.2 costs $8 per million input and $40 per million output tokens, against Astra's $10 and $50. The author's testing suggests Anthropic models use more tokens, so the lower list price did not yield a lower bill.
Anthropic has launched a self-hosted version of Claude Managed Agents in private beta, letting customers run both the agent coordination server and tool sandboxes in their own cloud account or on-premises data center. Credentials and session histories stay on customer infrastructure, while Claude requests go to a configured inference provider. The license requires responding to safety notices within 8 hours, limits which models can be used, and bars blocking certain data from being sent to Anthropic. The hosted service debuted in April, followed by self-hosted sandboxes the next month.
Elastic search executive Sri Desikan told The New Stack that a context gap is keeping enterprises from getting expected returns on AI agent investments. He argues that better frontier models waste token spending if not fed proper context, and that simply connecting a corpus via MCP servers or loading all data into a context window is computationally expensive. Delivering the right information to agents quickly is framed as the path from demos to durable automation.
DeepSeek's open-source coding agent, DeepSeek Harness, ranks sixth on Terminal-Bench even though it is statistically tied for first across the five scored evaluations on a leaderboard run by Hugging Face researcher Joël Niklaus. The leaderboard compares ten harnesses running the same models. The agent shuts down background processes when a task finishes, which breaks three Terminal-Bench tasks that need those processes to keep running. It passed just two of 36 attempts on those tasks using DeepSeek-V4.1-Flash and GLM-5.3-Flash.
The New Stack is hosting a webinar with Vespa.ai's Jenny Morris on Wednesday, December 2, on preparing company data for agentic retrieval. The session argues that upgrading to a more capable AI model will not fix a failing agent workflow unless the underlying data is organised for fast, accurate search.
Google and Cursor both introduced features this week to stop AI coding agents from duplicating work and conflicting with each other's changes. Google's Swarm organizes multiple agents working for a single developer, while Cursor's multiplayer Projects lets agents belonging to different colleagues coordinate. Cursor first launched Projects in September, using a coordinator agent that splits a larger job into tasks for worker agents. Cursor was acquired by SpaceX for $60 billion in August.
Chainguard CTO Matt Moore defended at KubeCon + CloudNativeCon North America 2026 the company's release of patched PHP container images more than six hours before the October 21 coordinated disclosure lifted. PHP maintainers accused Chainguard of giving attackers a head start, but Moore said the company believed attackers could already exploit, or were close to exploiting, CVE-2026-109618 and CVE-2026-109619, citing security vendors' reports of activity. "We own the decision," he said.